MCP control plane for cloud subagents

Spawn a cloud agent without trapping your lead agent in a long-running task.

AgentDispatch gives OpenClaw, Hermes, Claude Code, Codex, and custom orchestrators one provider-neutral tool: hand off the task, keep durable status, and continue through A2A or another runtime protocol when the cloud supports it.

Start with AWS AgentCore View MCP server

1 tool: spawn_cloud_agent
V1 runtime: AWS AgentCore
Future path: GCP, Azure, Kubernetes, local

spawn_cloud_agent({
  runtime: "research-agent",
  instruction: "Audit this repo in the background.",
  protocol: "a2a",
  context: {
    repo: "agent-dispatch",
    priority: "background"
  }
})

→ {
  task_id: "task_...",
  status: "running",
  cloud_agent: {
    protocol: "a2a",
    invocation: { runtimeUrl: "https://..." },
    a2a: { messageMethod: "message/send" }
  }
}

Architecture

Control plane stays stable. Runtime adapters do the cloud-specific work.

Lead agent calls MCP

The local agent calls spawn_cloud_agent or the lower-level dispatch_task tool.

Core routes the request

AgentDispatch resolves provider, account profile, capability, task type, target mode, and runtime protocol.

Adapter starts the runtime

AWS AgentCore V1 invokes an existing runtime session or provisions a per-task runtime from an ECR image.

Agent keeps interacting

The response returns durable polling tools plus A2A, MCP, AG-UI, or HTTP metadata when available.

Why it matters

Local agents should orchestrate. Cloud agents should execute.

Long-running work: deep research, audits, migrations, batch analysis, and build pipelines do not block the local session.

Credential boundaries: cloud credentials stay in account profiles and provider SDK chains, never inside MCP tool payloads.

Provider expansion: new providers are adapter packages, not new MCP tools that every lead agent must learn.

Package map

Separate repos, one contract.

@agent-dispatch/coreRuntime models, routing, adapter contract. @agent-dispatch/mcp-serverMCP tools for spawn, status, logs, results, cancel. @agent-dispatch/adapter-aws-agentcoreAWS AgentCore runtime adapter. @agent-dispatch/worker-agentcoreReference AgentCore worker with A2A endpoint. @agent-dispatch/sdkTyped client for apps and frameworks. @agent-dispatch/cliConfig bootstrap and diagnostics. @agent-dispatch/store-sqliteLocal durable task and event store. @agent-dispatch/adapter-templateStarting point for future clouds.

V1 implementation

AWS AgentCore runtime first.

The first adapter supports agent.run and command.run, session mode for existing runtime ARNs, runtime mode for ECR-backed per-task runtimes, event normalization, cancellation, cleanup, and A2A handoff metadata.

Runtime protocol hints for A2A, MCP, AG-UI, and HTTP.
Named account profiles using standard AWS credential chains.
SQLite task state, logs, provider refs, and artifact records.
Structured clarifications when a cloud spawn needs missing inputs.

Quickstart

Configure once. Spawn when needed.

npm install -g @agent-dispatch/cli

agentdispatch init \
  --region us-west-2 \
  --runtime-arn arn:aws:bedrock-agentcore:... \
  --protocol a2a

agentdispatch doctor

{
  "mcpServers": {
    "agentdispatch": {
      "command": "npx",
      "args": [
        "-y",
        "@agent-dispatch/mcp-server",
        "--config",
        "/absolute/path/agentdispatch.config.json"
      ]
    }
  }
}

Roadmap

Cloud-neutral from day one.

AWSAgentCore, ECS, App Runner, Lambda, Batch.

GCPCloud Run services and jobs, Vertex AI Agent Engine.

AzureContainer Apps, App Service, AI Foundry runtimes.

KubernetesDeployments, Jobs, CronJobs, service exposure.